work / gameloom · r&d · 2026

GameLoom

Type a prompt, get a working mobile game

proof26 complete apps generated end to end

harnesscustom file / project / dependency / server tools

modelsmulti-provider routing: Anthropic, OpenAI, Google

feedbackSSE streams every tool call to the browser live

system

  prompt ──▸ planner ──▸ tool-calling loop
                            │
              ┌─────────────┼──────────────┐
              ▼             ▼              ▼
         file tools   dependency      server tools
         (write src)  tools (install) (run preview)
              │
              ▼
     SSE stream ──▸ browser: live build log + playable preview

the experiment

Could 'Lovable for mobile games' work? The answer required building a real coding agent: not a prompt template, but a harness where a model plans, writes files, installs dependencies, starts servers, and recovers when a step fails.

how it works

The backend exposes typed tools — file operations, project scaffolding, dependency management, dev-server control — and routes between Anthropic, OpenAI, and Google models per task. Every tool call streams to the browser over SSE, so you watch the game assemble itself, then play it in a live preview.

The repo contains 26 generated apps: the difference between a demo that worked once and a system that works.

what it taught me

Agent harness design is product design. Tool granularity, error recovery, and what the model is allowed to see determine whether generation converges or spirals. That lesson transfers directly to any 'AI that does real work' system I build for clients.

stack

Vercel AI SDK · Next.js 15 · Express + TS · Drizzle + Postgres · Expo / React Native

Private codebase.

gameloom.app (live)